Application of deep learning techniques for exoplanet detection in high contrast imaging

The direct imaging of exoplanets through 10-m class ground-based telescopes
is now a reality of modern astrophysics. Reaching this milestone is the re-
sult of significant advances in the field of high-contrast imaging, marked by
the development of dedicated telescope instrumentation, including extreme
adaptive optics and advanced coronagraphy. However, despite these advance-
ments, residual optical aberrations persist, generating quasi-static speckles in
the image field of view, whose shape and intensity are similar to potential
companions. Over the past two decades, numerous image post-processing
techniques have been proposed to further eliminate this residual speckle
noise and detect the planet signature. Among these techniques, supervised
deep learning was introduced through the SODINN detection algorithm, a
binary classifier that uses a convolutional neural network to learn a model
that distinguishes between noise and the point-like source in the image.
The recent Exoplanet Imaging Data Challenge (EIDC) served as a platform for
benchmarking SODINN and other image post-processing techniques. The
results revealed that SODINN tends to produce a notable number of false
positives, while the most effective techniques rely on mechanisms to capture
local image noise dependencies. Building upon these findings, this thesis
aims to improve the detection performance of SODINN through capturing
new local noise dependencies. Through the development of new statistical
methods, we explore the possibility to identify different noise regimes across
the angular differential imaging processed image and adapt the SODINN
neural network, and its training process, to work under this stratification
strategy. This model adaptation leads to the creation of a new detection algo-
rithm, called NA-SODINN. Through ROC analyses, NA-SODINN undergoes
rigorous testing against its predecessor, demonstrating an improved balance
between sensitivity and specificity in detection. Furthermore, NA-SODINN
is benchmarked against the full set of detection algorithms submitted in
the EIDC. The results indicate that NA-SODINN either matches or exceeds
the performance of the most powerful detection algorithms. NA-SODINN
is finally used to reanalyze a filtered sample from the recent SHINE exo-
planet survey, providing valuable insights and potential exoplanet candidates.
Throughout the supervised machine learning case, this study illustrates and
reinforces the importance of adapting the task of detection to the local content
of processed images.